Sprinter1999 / FedLSR

Official codes for CIKM '22 full paper: Towards Federated Learning against Noisy Labels via Local Self-Regularization

FedLSR

This repository is the official Pytorch implementation DEMO of FedLSR framework in this paper:

Towards Federated Learning against Noisy Labels via Local Self-Regularization (CIKM 2022 full paper track)

The slides for the oral presentation link is here.

Note: For researchers who are interested to adopt this as a baseline, please take care of the difference between FedLSR and FedLSR+ (which is given in the discussion section). The key motivation to append an entropy regularization term (to devise FedLSR+) is to further make the model more robust to extreme noisy labels. For experiments on clothing1M, it is suggested to raise the learning rate to 0.1 for FedLSR's implementation.

---

Main Baselines

• FedAvg [paper]
• Symmetric Cross Entropy [paper] [code]
• Co-teaching [paper] [code]
• Robust Federated Learning [paper] [code]

---

Requirements

• Python: 3.8
• Pytorch: 1.7.1
• torchvision: 0.8.2
• Other dependencies

---

Special Thanks

• Girum & IAMjmj give some valuable comments in the Github Issue part, and I just clarify some missed points of this paper. Please visit there for more information.

---

Citing this work

 @inproceedings{jiang2022towards,
  title={Towards federated learning against noisy labels via local self-regularization},
  author={Jiang, Xuefeng and Sun, Sheng and Wang, Yuwei and Liu, Min},
  booktitle={Proceedings of the 31st ACM International Conference on Information \& Knowledge Management},
  pages={862--873},
  year={2022}
}